Language Log

Sean Hannity's new book, Live Free or Die, was released a couple of days ago. The original cover featured a Latin motto, "Vivamus vel libero perit Americae", whose source was apparently Google Translate's version of "Live Free or America Dies":

As Spencer Alexander McDaniel observed, this is gobbledygook — or perhaps we should say "googledygook".

The title of McDaniel's post ("Sean Hannity does not know Latin") is unfair, though probably true, since the cover design was most likely created by the publisher. But a better observation might have been that Google Translate doesn't know Latin. Of course modern machine translation systems don't really "know" any languages, but have just memorized patterns of contextual correspondences. When the available training data is merely a few million words, as in the case of Latin, the results are often bad, as here.

Today's adventure in AI brought yet another robocall, which my Google Assistant intercepted since the calling number (probably spoofed) was not in my contacts list. Here's Google Assistant's rendering of the interaction

News sources around the world reported recently on a tragedy — "Officials: Chinese Ambassador to Israel Found Dead in Home", Associated Press 5/17/2020:

JERUSALEM — The Chinese ambassador to Israel was found dead in his home north of Tel Aviv on Sunday, Israel’s Foreign Ministry said.

Israeli Police Spokesman Micky Rosenfeld said the ambassador's death was believed to be from natural causes.

Du Wei, 58, was appointed envoy in February in the midst of the coronavirus pandemic. He previously served as China’s envoy to Ukraine. He was found dead at the ambassador's official residence in Herzliya.

But if you read it on Israel Hayom (or other Hebrew sources) via Google Translate, you'll get a different picture:

Today's xkcd:

Email today from Edwin Williams:

I constructed "new" sonnets from Shakespeare's sonnets by this formula: from a set of 7 randomly selected Shakespeare sonnets (a…g) I made a new sonnet "a b a b c d c d e f e f g g", which means, the first line is taken from the first line of sonnet a, the second line from sonnet the second line of sonnet b, etc. So no two adjacent lines were from the same sonnet, except the last two. I made 154 of these (same number as S made).

I did it for fun but was startled by the result–the new sonnets were sonnetlike, felt syntactically coherent, and begged for interpretation. People I sent them to were fascinated by them, even when they saw what I had done, or after I told them. One of them, Craig Dworkin, a poet I got to know when he was at Princeton in the 90s, asked to include them in his e-poetry site, and there they sit: http://eclipsearchive.org/projects/EDWIN/.

This is another note on the amazing ability of modern AI learning techniques to imitate some aspects of natural-language patterning almost perfectly, while managing to miss common sense almost entirely. This probably tells us something about modern AI and also about language, though we probably won't understand what it's telling us until many years in the future.

Today's example comes from Zihang Da et al., "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", arXiv 6/2/2019.

Following up "Shelties On Alki Story Forest" (11/26/2019) and "The right boot of the warner of the baron" (12/6/2019), here's some recent testimony from engineers at Google about the brittleness of contemporary speech-to-text systems: Arun Narayanan et al., "Recognizing Long-Form Speech Using Streaming End-To-End Models", arXiv 10/24/2019.

The goal of that paper is to document some methods for making things better. But I want to underline the fact that considerable headroom remains, even with the massive amounts of training material and computational resources available to a company like Google.

Modern AI (almost) works because of machine learning techniques that find patterns in training data, rather than relying on human programming of explicit rules. A weakness of this approach has always been that generalization to material different in any way from the training set can be unpredictably poor. (Though of course rule- or constraint-based approaches to AI generally never even got off the ground at all.) "End-to-end"  techniques, which eliminate human-defined layers like words, so that speech-to-text systems learn to map directly between sound waveforms and letter strings, are especially brittle.

Oscar Schwartz, "Natural Language Processing Dates Back to Kabbalist Mystics", IEEE Spectrum 10/28/2019 ("Long before NLP became a hot field in AI, people devised rules and machines to manipulate language"):

The story begins in medieval Spain. In the late 1200s, a Jewish mystic by the name of Abraham Abulafia sat down at a table in his small house in Barcelona, picked up a quill, dipped it in ink, and began combining the letters of the Hebrew alphabet in strange and seemingly random ways. Aleph with Bet, Bet with Gimmel, Gimmel with Aleph and Bet, and so on.

Abulafia called this practice “the science of the combination of letters.” He wasn’t actually combining letters at random; instead he was carefully following a secret set of rules that he had devised while studying an ancient Kabbalistic text called the Sefer Yetsirah. This book describes how God created “all that is formed and all that is spoken” by combining Hebrew letters according to sacred formulas. In one section, God exhausts all possible two-letter combinations of the 22 Hebrew letters.

By studying the Sefer Yetsirah, Abulafia gained the insight that linguistic symbols can be manipulated with formal rules in order to create new, interesting, insightful sentences. To this end, he spent months generating thousands of combinations of the 22 letters of the Hebrew alphabet and eventually emerged with a series of books that he claimed were endowed with prophetic wisdom.

Yiming Wang et al., "Espresso: A fast end-to-end neural speech recognition toolkit", ASRU 2019:

We present ESPRESSO, an open-source, modular, extensible end-to-end neural automatic speech recognition (ASR) toolkit based on the deep learning library PyTorch and the popular neural machine translation toolkit FAIRSEQ. ESPRESSO supports distributed training across GPUs and computing nodes, and features various decoding approaches commonly employed in ASR, including look-ahead word-based language model fusion, for which a fast, parallelized decoder is implemented. ESPRESSO achieves state-of-the-art ASR performance on the WSJ, LibriSpeech, and Switchboard data sets among other end-to-end systems without data augmentation, and is 4–11× faster for decoding than similar systems (e.g. ESPNET)

Today's SMBC:

Mouseover title: "The other day I was really freaked out that a computer could generate faces of people who DON'T REALLY EXIST, only to later realize painters have been doing this for several millenia."

A couple of years ago, in connection with the JSALT2017 summer workshop, I tried several commercial speech-to-text APIs on some clinical recordings, with very poor results. Recently I thought I'd try again, to see how things have progressed. After all, there have been recent claims of "human parity" in various speech-to-text applications, and (for example) Google's Cloud Speech-to-Text tells us that it will "Apply the most advanced deep-learning neural network algorithms to audio for speech recognition with unparalleled accuracy", and that "Cloud Speech-to-Text accuracy improves over time as Google improves the internal speech recognition technology used by Google products."

So I picked one of the better-quality recordings of neuropsychological test sessions that we analyzed during that 2017 workshop, and tried a few segments. Executive summary: general human parity in automatic speech-to-text is still a ways off, at least for inputs like these.

At least I think that's what happened here. Gita Jackson, "Tumblr's New Algorithm Thinks Garfield Is Explicit Content", Kotaku 12/4/2018:

Yesterday, Tumblr announced that it will ban all adult content starting December 17th. As users logged into their accounts, they have seen that some of their posts now have a red banner across them, marking them as flagged for explicit content. The problem is, a lot of these posts are hilariously far from being pornographic.

It’s pretty clear that these flags are being done based on an algorithm, and the algorithm is finding false positives. Here’s a list of things that got flagged: a fully clothed woman, a drawing of a dragon, fan-art of of characters from the anime Haikyu!!, art from the children’s book The Princess Who Saved Herself that the author of said book posted, a drawing of a bowl of fruit with mouths, a video of abstract blurs, Garfield.

Just checking to see that Google Translate is still into hallucinatory automatic writing…

Today's input is five random hiragana characters — あっぉぉを — repeated various numbers of times: